Kling 3.0 First-Last-Frame to Video

You give Kling V3 a first frame and a last frame, and it invents the whole movie between them - the movement, the physics, the in-between acting that turns two stills into a scene. That's the entire job of this node, and it's the same bookend idea as the v2.5-era "Kling Start-End Frame" node, just on Kling's newer V3 model with audio generation on by default. If your clip needs to end at a specific image - a character arriving at a destination, a door closing, a cut that lands exactly where the next scene needs it - this is the built-in for it.

It's an API node like the rest of the partner/video family: ships with ComfyUI core, uploads both frames to Comfy Org's backend, calls Kling's image2video endpoint with the second frame as the image_tail, polls until done, and hands back a VIDEO. Nothing renders locally; it's billed to your Comfy account in credits.

The inputs that matter

• first_frame and end_frame - the bookends. Both must be at least 300×300 with aspect ratios between 1:2.5 and 2.5:1. The less your two frames agree on subject, lighting, and camera height, the harder the model works and the more visible the invention gets.
• duration - 3 to 15 seconds, default 5. Give a far-apart pair of frames more room.
• model - a dynamic combo: kling-v3, and inside it a resolution pick of 720p, 1080p, or 4K. That resolution choice is where the real cost lives - 4K is dramatically pricier than 720p, and the node maps it to Kling's std/pro/4k modes internally.
• generate_audio - default on for this node, which is a genuine feature: V3 can produce synchronized sound for the clip. It's also a cost multiplier, so if you don't need it, flip it off.
• prompt - what happens between the frames, described in words. seed - the family standard: a re-run toggle, non-deterministic results regardless.

Output is a single VIDEO.

Where people get burned

• 4K creep. The resolution is buried in the model combo, so it's easy to leave it on 4K and watch the price per second multiply. Check it before you queue.
• Bookends that disagree. A first frame and a last frame that share almost nothing - different person, different lighting - force the model to hallucinate the bridge, and it shows in exactly the way you'd expect. Design the pair as one scene, not two unrelated images.
• Audio on by default. It's a pleasant surprise until you're iterating 20 takes of a silent test clip and wondering why each one costs more than the badge on the other Kling nodes.
• Still nondeterministic. The seed won't reproduce a result; it only forces a new draw.

If you're building sequences that need precise landing frames, this node plus the v3 quality is the right combination - just treat the model resolution pick and the audio toggle as the budget controls they are.